OpenMP, gfortran and the curse of non-speedup

Question

I have a ridiculously parallelizable (in my mind) piece of Fortran code, so I thought I could give OpenMP a try and see if any speedup could be achieved. This is on Windows 10, MinGW64 with gcc 11.2. I have 16 cores and 256 GB of RAM on an Intel Xeon Gold 5122 @ 3.6 GHz.

The code is definitely not a big deal:

  start_time = omp_get_wtime() 
  !$omp parallel do default(shared) private(i)
  do i = 1, nF   
    call something(input1, input2(i), input3, input4, output1(i, :), output2(i, :), output3(i, :))   
  end do 
  !$omp end parallel do
  end_time = omp_get_wtime()       
  
  write(*, *) 'Number of threads:', omp_get_max_threads(), ', Elapsed time:', end_time-start_time

nF is in the order of 8,600. The calculations are completely independent on which i index I am looping on, it doesn't matter in which order they are executed. There is no interaction between the inputs or the outputs at iteration i-1 with their values at iteration i. Whatever goes on inside the subroutine something is pure calculations, and not that tough either.

Now, I am not expecting magical linear speed improvements, but this is what I get:

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=1 & test.exe
 Number of threads:           1 , Elapsed time:  0.43099999427795410

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=2 & test.exe
 Number of threads:           2 , Elapsed time:  0.75899982452392578

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=4 & test.exe
 Number of threads:           4 , Elapsed time:  0.69499993324279785

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=8 & test.exe
 Number of threads:           8 , Elapsed time:  0.55299997329711914

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=16 & test.exe
 Number of threads:          16 , Elapsed time:  0.57800006866455078

Happily, no speed up. And actually a slowdown, no matter how many threads I tell OpenMP to use. It may be that for a code that takes 0.43 seconds to run it's pointless to attempt to use OpenMP, as the thread creation time will always swamp whatever improvement they may bring. It may also be I am misunderstanding how all of this work.

Optimization switches I use to compile the code:

-O3 -funroll-loops -march=native -fno-asynchronous-unwind-tables -fopenmp

Any explanation is most welcome, and suggestions on modifying my !$omp stuff is also welcome.

EDIT

I have edited the code by declaring the matrices outout1, output2 etc... as "transposed", so it looks like this:

  start_time = omp_get_wtime() 
  !$omp parallel do default(shared) private(i)
  do i = 1, nF   
    call something(input1, input2(i), input3, input4, output1(:, i), output2(:, i), output3(:, i))   
  end do 
  !$omp end parallel do
  end_time = omp_get_wtime()       
  
  write(*, *) 'Number of threads:', omp_get_max_threads(), ', Elapsed time:', end_time-start_time

No changes in timings (it's possibly even worse):

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=1 & test.exe
 Number of threads:           1 , Elapsed time:  0.43000006675720215

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=2 & test.exe
 Number of threads:           2 , Elapsed time:  0.63800001144409180

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=4 & test.exe
 Number of threads:           4 , Elapsed time:  0.96600008010864258

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=8 & test.exe
 Number of threads:           8 , Elapsed time:  0.80500006675720215

C:\Users\User\MyProjects\Test>set OMP_NUM_THREADS=16 & test.exe
 Number of threads:          16 , Elapsed time:  0.75200009346008301

OpenMP, gfortran and the curse of non-speedup

0 Answers